The radiotherapy planning CT-based multi-omics for predicting the radiation pneumonitis in lung cancer patients: A multi-center study

Abstract

Abstract Background To predict the risk of radiation pneumonitis (RP), deep learning (DL) models were built to stratify lung cancer patients. Our study also investigated the impact of RP on survival. Methods This study retrospectively collected 100 RP and 99 matched non-RP lung cancer patients treated with radiotherapy from two independent centers. These patients were randomly divided into training (n = 175) and validation cohorts (n = 24). The radiomics and dosiomics features were extracted from radiation planning computed tomography (CT). Clinical information was retrospectively collected from the electronic medical record database. All features were screened by LASSO cox regression. A multi-omics prediction model was developed by the optimal algorithm and estimated the area under the receiver operating characteristic curve (AUC). Overall survival (OS) between RP, non-RP, mild-RP, and severe-RP groups was analyzed by the Kaplan-Meier method. Results There were eventually selected 16 radiomics features, 2 dosiomics features, and 1 clinical feature to build the best multi-omics model. GLRLM_Gray Level Non Uniformity Normalized and GLCM_MCC from PTV were essential dosiomics features, and T stage was a paramount clinical feature. The optimal performance for predicting RP was the AUC of testing set [0.94, 95% confidence interval (CI) (0.939-1.000)] and the AUC of external validation set [0.92, 95% CI (0.80-1.00)]. All RP patients were divided into mild-RP and severe-RP group according to RP grade (≤ 2 grade and > 2 grade). The median OS was 31 months (95% CI, 28–39) for non-RP group compared with 49 months (95% CI, 36-NA) for RP group (HR = 0.53, P = 0.0022). Among RP subgroup, the median OS was 57months (95% CI, 47-NA) for mild-RP and 25 months (95% CI, 29-NA) for severe-RP, and mild-RP group exhibited a longer OS (HR = 3.72, P < 0.0001). Conclusion The multi-omics model contributed to improvement in the accuracy of the RP prediction. Interestingly, this study also demonstrated that compared with non-RP patients, RP patients displayed longer OS, especially mild-RP.